Hiring someone to pull data out of PDFs and invoices with AI: what to specify and how to test it

The short answer

Pulling data out of PDFs and invoices with AI is easy to demo and hard to trust. A demo shows three clean invoices turned into three tidy rows. Your real inbox has scans, credit notes, two-page invoices, the same invoice sent twice and a supplier who changed their template last month. The difference between those two worlds is where the money goes wrong.

So brief three things before anyone quotes. Which fields, exactly, with the format each must arrive in. The messy cases, named, with what should happen to each. And how accuracy will be measured: a set of your own documents with the correct answers written down, checked field by field. Everything else (which tool, which AI model, which platform) the freelancer can propose. The workflow automation brief gives you the general fields; this guide is about what to write in them for documents.

Why extraction goes wrong quietly

A broken email automation is visible: the email does not arrive. A broken extraction is not. A row appears, every column is filled, and one of them is wrong. The total includes tax when it should not. The date is read as day-month instead of month-day. The supplier’s bank details are taken from the footer of a different company. Nobody notices until a payment or a report is off.

I learned the general shape of this from a data collector of my own, not from invoices. It read reviews from a set of marketplace pages, ran without a single error, and produced a count that looked plausible. The real number was 317, not 437: the page had loaded some entries twice and the collector kept both copies. The full note on that recount explains how it was found. The lesson carries straight over to documents: an extraction job that looks correct from the inside needs something outside it to check against.

Three kinds of error cover most of what goes wrong:

• Wrong value. The field is filled, but with the wrong number or text: a subtotal instead of a total, the issue date instead of the due date.
• Missing or invented value. The field is blank when the document has it, or, worse with AI, filled with something the document does not say.
• Wrong count. Duplicates kept, a two-page invoice split into two, a credit note treated as an invoice.

What to put in the brief

The general brief fields still apply. Here is what to write in each when the input is documents:

Two of these deserve a closer look.

The field table is the contract. “Extract the invoice data” is not a specification. A table of twelve named fields with an example value each is. It also makes the quote comparable between sellers, because each one is pricing the same twelve fields.

The unreadable path is a design decision. One of my own scheduled jobs reads files when it runs unattended. On one run, the connection to the document-reading tool it relied on timed out, and it could not read the file at all. A job that could fill fields anyway would have produced a confident, empty or invented row. Decide in the brief what happens instead, and test it.

Build a test set before you hire

This is the part most buyers skip, and it is the part that makes everything else work. Before you message anyone, collect about 30 real documents and write the correct answer for every field in a spreadsheet. Then make sure the set includes the cases that break things:

Include	Why
3–5 scans or phone photos, one of them poor	Reading quality drops fastest here
2 invoices from the same supplier on different templates	Catches extraction that depends on layout
A multi-page invoice	Totals often sit on the last page
A credit note	Should not be counted as a cost
The same invoice twice	Should produce one row, not two
An invoice in another currency or language, if you get any	Formats and decimal separators differ
One file that is not an invoice at all	Should be set aside, not extracted

Keep the answer sheet to yourself until delivery. Give the freelancer a separate sample of documents to build with. If they build and test on the same 30 you will accept against, the test measures memory, not accuracy.

Thirty documents sounds small, and it is. It is enough to see whether one field is systematically wrong, which is what you need to know before paying. For a bigger volume, ask for the same test to be repeated on a fresh sample after the first month.

How to accept it

Run the test set yourself, or watch the freelancer run it on a call, and compare every field with your answer sheet.

Point 4 matters more than it looks. In a scheduled batch job of mine, three items wrote one field in the wrong format. They broke a step the whole run shared, and the three batches after them, about 15 items that had passed their own check, stopped moving. Nothing told me; I found it by reading the job log. The note on that batch is the reason I now ask what happens to the rest of a batch when some items are bad. For invoices, the answer should be: the bad one waits for a person, the good ones carry on.

What moves the price

On the workflow automation job page, the basic packages of the vetted sellers started at $100 to $600 when checked in September 2026, and a basic package usually covers one trigger and one or two connected tools. Document extraction with an AI step is quoted above the basic tier. What moves it:

• How varied the documents are. Ten suppliers on clean digital PDFs is a different job from two hundred suppliers with scans.
• The number of fields and rules. Line items (every row of the invoice table) cost far more than header fields (supplier, date, total).
• Where the output goes. A spreadsheet is simpler than writing into an accounting tool through its API.
• The review step. A screen where a person approves uncertain rows before they are posted adds build time and saves mistakes.
• Volume. AI usage is billed per document read. Ask for a cost per 100 documents on your account; the running costs guide lists the other monthly lines.

If the documents feed a chatbot that answers questions about them, rather than rows in a sheet, it is a different order: see the chatbot job page. Sellers who build around documents for that purpose, such as Bensbahou, describe answering from your own documents and data in their gigs.

Where it goes wrong anyway

• A supplier changes their template. Accuracy drops for that supplier only. A monthly spot check of five rows catches it; the fix belongs in the maintenance terms, not a new negotiation.
• Someone forwards the same invoice twice. If the duplicate rule was never tested, you pay twice. The idempotency explainer covers why “run it again” must not mean “post it again”.
• The AI fills a gap. A due date that is not on the invoice appears anyway. That is why “nothing invented” is an acceptance line, and why the rules against made-up answers apply to extraction too.

Wrap-up

An extraction job is only as good as the test it passed. Write the field table, name the messy cases, decide what happens to a document it cannot read, and keep a 30-document answer sheet of your own. Then the quote is comparable, the acceptance is a count rather than an impression, and the first wrong total is caught by your spot check instead of your accountant.